fpls-08-00770 May 12, 2017 Time: 17:38 # 1
ORIGINAL RESEARCHpublished: 15 May 2017
doi: 10.3389/fpls.2017.00770
Edited by:Susan L. Ustin,
University of California, Davis, USA
Reviewed by:Kyla Dahlin,
Michigan State University, USAE. Natasha Stavros,
Jet Propulsion Laboratory, USA
*Correspondence:Tuyet T. A. Truong
Specialty section:This article was submitted to
Functional Plant Ecology,a section of the journal
Frontiers in Plant Science
Received: 28 October 2016Accepted: 25 April 2017Published: 15 May 2017
Citation:Truong TTA, Hardy GESJ and
Andrew ME (2017) ContemporaryRemotely Sensed Data Products
Refine Invasive Plants Risk Mappingin Data Poor Regions.
Front. Plant Sci. 8:770.doi: 10.3389/fpls.2017.00770
Contemporary Remotely SensedData Products Refine Invasive PlantsRisk Mapping in Data Poor RegionsTuyet T. A. Truong1,2*, Giles E. St. J. Hardy1 and Margaret E. Andrew1
1 Environmental and Conservation Sciences, School of Veterinary and Life Sciences, Murdoch University, Perth, WA,Australia, 2 Faculty of Environment, Thai Nguyen University of Agriculture and Forestry, Thai Nguyen, Vietnam
Invasive weeds are a serious problem worldwide, threatening biodiversity and damagingeconomies. Modeling potential distributions of invasive weeds can prioritize locations formonitoring and control efforts, increasing management efficiency. Forecasts of invasionrisk at regional to continental scales are enabled by readily available downscaledclimate surfaces together with an increasing number of digitized and georeferencedspecies occurrence records and species distribution modeling techniques. However,predictions at a finer scale and in landscapes with less topographic variation may requirepredictors that capture biotic processes and local abiotic conditions. Contemporaryremote sensing (RS) data can enhance predictions by providing a range of spatialenvironmental data products at fine scale beyond climatic variables only. In this study,we used the Global Biodiversity Information Facility (GBIF) and empirical maximumentropy (MaxEnt) models to model the potential distributions of 14 invasive plantspecies across Southeast Asia (SEA), selected from regional and Vietnam’s lists ofpriority weeds. Spatial environmental variables used to map invasion risk includedbioclimatic layers and recent representations of global land cover, vegetation productivity(GPP), and soil properties developed from Earth observation data. Results showed thatcombining climate and RS data reduced predicted areas of suitable habitat comparedwith models using climate or RS data only, with no loss in model accuracy. However,contributions of RS variables were relatively limited, in part due to uncertainties inthe land cover data. We strongly encourage greater adoption of quantitative remotelysensed estimates of ecosystem structure and function for habitat suitability modeling.Through comprehensive maps of overall predicted area and diversity of invasive species,we found that among lifeforms (herb, shrub, and vine), shrub species have higherpotential invasion risk in SEA. Native invasive species, which are often overlooked inweed risk assessment, may be as serious a problem as non-native invasive species.Awareness of invasive weeds and their environmental impacts is still nascent in SEA andinformation is scarce. Freely available global spatial datasets, not least those providedby Earth observation programs, and the results of studies such as this one providecritical information that enables strategic management of environmental threats such asinvasive species.
Keywords: non-native invasive species, invasibility, MaxEnt, MODIS, native invasive species, species distributionmodeling, Southeast Asia
Frontiers in Plant Science | www.frontiersin.org 1 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/Plant_Science/editorialboardhttp://www.frontiersin.org/Plant_Science/editorialboardhttps://doi.org/10.3389/fpls.2017.00770http://creativecommons.org/licenses/by/4.0/https://doi.org/10.3389/fpls.2017.00770http://crossmark.crossref.org/dialog/?doi=10.3389/fpls.2017.00770&domain=pdf&date_stamp=2017-05-15http://journal.frontiersin.org/article/10.3389/fpls.2017.00770/abstracthttp://loop.frontiersin.org/people/388073/overviewhttp://loop.frontiersin.org/people/410577/overviewhttp://loop.frontiersin.org/people/349683/overviewhttp://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 2
Truong et al. Weed Risk Mapping in Southeast Asia
INTRODUCTION
Invasive plants have emerged as a serious problem for globalbiodiversity. Their infestations can lead to the extinction(Groves et al., 2003) and endangerment (Wilcove et al., 1998;Pimentel et al., 2005) of native species and the alteration ofecosystem processes (Vitousek and Walker, 1989; Simberloff,2000). Although invasive species that are introduced to a regionreceive the greatest attention, it is not necessary for a species tobe non-native to be invasive. Invasive species are defined as thosethat are expanding their range (Valéry et al., 2008). Under globalclimate change and human disturbance, some native species havealso become aggressive invasive weeds (Avril and Kelty, 1999;Wang et al., 2005; Hooftman et al., 2006; Valéry et al., 2009; Leet al., 2012). Given the large impacts that invasive species can haveand the limited possibilities for eradication, early detection andprevention of the establishment of invasive species should be apriority in conservation policies (Genovesi, 2005). Identificationof areas that are at potential invasion risk, to either non-native ornative invasive species, can be an effective way to guide efficientmanagement and prevent further incursion (Kulhanek et al.,2011).
Species distribution models (SDMs) are currently a popularmethod for predicting the geographic distribution of species(Peterson, 2006). They are developed statistically from theknown occurrences of the species and characteristics of theenvironment to identify similar suitable habitats and, thereby,predict the geographic distribution in unknown regions (Guisanand Zimmermann, 2000; Peterson and Vieglais, 2001; Peterson,2006; Pearson, 2010). Given these modest data requirements, theyare especially useful in cases of poorly studied taxa (Kearney andPorter, 2009). Therefore, SDMs have become an important toolto investigations of invasibility that aim to predict the potentialdistributions of invasive species (Peterson, 2003; Thuiller, 2005).Since the early study of Peterson et al. (2003) in predicting thepotential distribution of four invasive plants in North America,SDMs have been increasingly and widely applied all over theworld to predict biological invasions (Guisan and Thuiller,2005; Underwood et al., 2013), especially exotic plants (Zhuet al., 2007; Andrew and Ustin, 2009; Barik and Adhikari, 2011;Fernández et al., 2012; Rameshprabu and Swamy, 2015). InSDMs, the environmental variables used vary at different scales(Bradley et al., 2012). At regional to continental scales, forecastsof invasion risk are often mainly driven by climatic factors(Pearson and Dawson, 2003). Predictions at a finer scale and inlandscapes with less topographic variation may require predictorsthat capture biotic processes (e.g., vegetation productivity) andlocal abiotic conditions (e.g., topography, soil type) (Pearson andDawson, 2003). However, continuous spatial measurements ofthese finer-scaled environmental variables are difficult to acquireat large spatial extent (Bradley et al., 2012).
Contemporary remote sensing (RS) now provides widelyavailable data products at multiple spatial and temporalresolutions that characterize a range of ecologically relevantpatterns and processes (Andrew et al., 2014). These data canbe used to measure habitat properties over a larger area thancan easily be covered by field surveys (Estes et al., 2008) and
augment the array of spatial environmental variables availableto SDMs to characterize abiotic and biotic niche axes beyondsimply climatic factors. Table 1 provides an overview of theremotely sensed information that has been incorporated intoSDMs as environmental predictor variables, to date, giving anindication of the evenness of research efforts and the capabilitiesof RS that are still relatively under-utilized. The most commonlyused variable extracted from RS data is topography/elevation(42% of 39 reviewed studies that have developed SDMs of plantspecies using RS predictors). Besides, other abiotic predictorshave been developed such as remotely sensed estimates of climateand weather, including surface temperature from sensors such asMODIS and rainfall estimates from TRMM and, more recently,the Global Precipitation Measurement mission, although studiesapplying these predictors are limited (Table 1). Soil properties,one of the most important factors for plant distributions andspecies invasion (Radosevich et al., 2007), is rarely studied (Heet al., 2015), although several recent studies have explored theuse of remotely sensed indicators of soil characteristics in SDMs(Table 1).
In addition to abiotic properties of the environment, bioticcharacteristics also play an important role in shaping species’spatial patterns (Wisz et al., 2013). RS can estimate manyproperties of the vegetated environment, and applications ofproducts such as land-cover data or vegetation proxies to SDMsare on the rise (Table 1). Land cover has been considered asthe primary determinant of species occurrences at a finer spatialresolution than climate (Pearson et al., 2004). Various studies(20% of 39 reviewed studies; Table 1) have applied land coverproducts derived from a variety of sensors (especially MODISand Landsat) to SDMs. However, most of the current landcover information is in categorical format, which can lead tothe propagation of classification errors (Cord and Rödder, 2011;Tuanmu and Jetz, 2014) and may not effectively represent theclasses most relevant to the species of interest. In contrast,remotely sensed estimates of continuously varying ecosystemproperties related to land cover and novel continuous land coverproducts can be used in SDMs and may avoid these limitations.Recent studies have found better performance from continuousestimates of vegetation properties and land cover rather thancategorical representations (Wilson et al., 2013; Cord et al., 2014b;Tuanmu and Jetz, 2014). A range of remotely sensed measures ofvegetation has been explored in SDMs, such as vegetation indices(Normalized difference vegetation index (NDVI), EnhancedVegetation Index), phenology, and canopy moisture in orderto evaluate variation in habitat quality at fine scales andin climatically homogenous regions (Table 1). Of vegetationmetrics, NDVI, a useful measure of vegetation properties, hasbeen extensively used as a predictor in SDMs (25.6%; Table 1).It represents photosynthetic activity and biomass in plants andis indirectly related to net primary production (Bradley andFleishman, 2008). However, a study of Phillips et al. (2008) notedthat while NDVI had high correlation with MODIS GPP (Grossprimary production) and NPP (Net primary production), it wasa less effective surrogate of productivity in areas of either sparseor dense vegetation. They found GPP to be better able to predictbiogeographic patterns of species richness (Phillips et al., 2008),
Frontiers in Plant Science | www.frontiersin.org 2 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 3
Truong et al. Weed Risk Mapping in Southeast Asia
but we know of no studies that have used GPP in SDMs. Value-added science products, such as the MODIS primary productivityproducts, may provide more meaningful depictions of vegetationprocesses and improved environmental predictor variables forspatial models of biodiversity (Phillips et al., 2008).
In addition to the typical niche axes used to inform variableselection for SDMs of plant species, there is a large body ofliterature determining the ecosystem properties that influenceinvasibility of a system, and these can be used to guideapplications of SDMs to evaluating invasion risk. Resourceavailability (e.g., light, CO2, water, nutrients) often facilitatessuccessful invasion. Invasibility is predicted to be greater in siteswith more unused resources (Davis et al., 2000). By damagingthe resident vegetation, disturbance reduces resource uptakeand competition, increasing resource availability (Hobbs, 1989;D’Antonio, 1993). Therefore, invasions by invasive plant speciesare often associated with disturbance (e.g., Fox and Fox, 1986).However, distributions of invasive species are typically modeledusing static environmental datasets that may poorly proxy thesedynamic processes (Franklin, 2010b; Dormann et al., 2012).Temporal summaries of GPP may provide useful indicators.GPP estimates total ecosystem photosynthesis, the cumulativeresponse of the vegetation to its environment, and may be used
as a spatial proxy of resource ability. As well, the variabilityof GPP over time can reflect disturbance processes (Goetzet al., 2012). Hence, quantitative spatial measurements of GPPare expected to be relevant predictor variables for modelinginvasibility. Also, including soil properties in SDMs may be usefulas numerous studies have shown that soil properties, includingnutrient availability, relate to invasibility (Huenneke et al., 1990;Burke, 1996; Harrison, 1999; Suding et al., 2004).
In this study, we hypothesize that the inclusion of recentlydeveloped global remotely sensed data products providingquantitative estimates of vegetation productivity and itsdynamics, land cover, and soil properties, in addition to climaticlayers, will enable a more complete representation of species’ecological niches by SDMs. To test the hypothesis, bioclimaticdata and RS data were used in isolated and combined modelspredicting the distribution of selected invasive plants acrossSoutheast Asia (SEA).
Southeast Asia is an important region to global biodiversity;it has four of the world’s 25 biodiversity hotspots (Sodhi et al.,2004). However, much biodiversity is being lost (Peh, 2010) dueto threatening processes such as habitat loss, degradation, climatechange, and pollution (Pallewatta et al., 2003). In addition,and operating in synergy with these anthropogenic changes,
TABLE 1 | Applications of remote sensing data as environmental variables in plant distribution models.
Predictor variables RS data source Reference
Abiotic predictors
Topographic data/elevation ASTER, Quickbird-2 andWorldView-2, LiDAR, SRTM
Rew, 2005; Bradley and Mustard, 2006; Buermann et al., 2008;Hoffman et al., 2008; Parviainen et al., 2008, 2013; Prates-Clark et al.,2008; Saatchi et al., 2008; Andrew and Ustin, 2009; Zellweger et al.,2013; Pottier et al., 2014; Pradervand et al., 2014; Questad et al.,2014; van Ewijk et al., 2014; Pouteau et al., 2015; Campos et al., 2016
Climate observations MODIS, TRMM, NASA Saatchi et al., 2008; Waltari et al., 2014; Deblauwe et al., 2016
Soil properties Landsat, MODIS Parviainen et al., 2013; Wang et al., 2016
Other physical variables (water, fire) MODIS, NASA Stohlgren et al., 2010; Cord and Rödder, 2011; Pau et al., 2013; Cordet al., 2014a
Land cover/land use MODIS, Landsat Pearson et al., 2004; Thuiller et al., 2004; Stohlgren et al., 2010;Morán-Ordóñez et al., 2012; Wilson et al., 2013; Cord et al., 2014b;Sousa-Silva et al., 2014; Tuanmu and Jetz, 2014; Gonçalves et al.,2016
Vegetation productivity
Normalized difference vegetation index (NDVI) Landsat, SPOT, MODIS Morisette et al., 2006; Zimmermann et al., 2007; Prates-Clark et al.,2008; Evangelista et al., 2009; Feilhauer et al., 2012; Engler et al.,2013; Parviainen et al., 2013; Schmidt et al., 2013; Zellweger et al.,2013; van Ewijk et al., 2014
Leaf area index (LAI) MODIS Buermann et al., 2008; Prates-Clark et al., 2008; Saatchi et al., 2008;Cord and Rödder, 2011; Engler et al., 2013
Enhanced Vegetation Index (EVI) MODIS Morisette et al., 2006; Stohlgren et al., 2010; Cord and Rödder, 2011;Schmidt et al., 2013; Cord et al., 2014a,b
Phenology MODIS, Landsat Bradley and Mustard, 2006; Morisette et al., 2006; Tuanmu et al., 2010;Gonçalves et al., 2016
Vegetation structure
Tree height LiDAR van Ewijk et al., 2014
Canopy roughness QSCAT Saatchi et al., 2008
Other vegetation properties
Canopy moisture Hyperspectral sensor, QSCAT Buermann et al., 2008; Prates-Clark et al., 2008
Spectral heterogeneity/functional types Hyperspectral sensor, Landsat Morán-Ordóñez et al., 2012; Schmidt et al., 2013; Henderson et al.,2014; Pottier et al., 2014
Frontiers in Plant Science | www.frontiersin.org 3 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 4
Truong et al. Weed Risk Mapping in Southeast Asia
invasive species damage the biodiversity and economy of theregion (Peh, 2010; Gower et al., 2012; Nghiem et al., 2013).Although impacts of invasive species in SEA are apparent,research on the level and types of impacts caused by invasivespecies is still limited (Nghiem et al., 2013). There are alsofew applications of SDM methods, either for invasive speciesor in general, in the region. Among studies about speciesdistributions worldwide, Porfirio et al. (2014) found only a smallfraction were conducted in Asia (∼3%). The absence of researchin this field is hindering SEA in providing a comprehensiveassessment of invasive species (Peh, 2010; Gower et al., 2012),and in effectively managing this aspect of global environmentalchange.
The goal of this study is to provide an overview ofpotential invasibility to 14 priority invasive plants in SEA. Togeneralize estimates of invasion risk across species traits that mayrequire different management approaches, we divided studiedspecies into different life forms (herb, vine, and shrub). Suchgroupings based on life-history attributes have been widelyused to understand the invasion process and propose tailoredmanagement strategies (McIntyre et al., 1995; Bear et al., 2006;Garrard et al., 2009). In addition, species were grouped bytheir origin status (native and non-native invasive species).Through evaluating SDMs by life forms and origin status, andusing different environmental predictor variable sets, our studyaddresses the following questions:
(i) Which life forms of invasive plant species pose the greatestrisk to SEA?
(ii) Are native weeds as great of a potential threat as non-native invasive species?
(iii) Do remotely sensed environmental predictor variablesimprove predictions of invasion risk over modelsconstructed with climate variables alone?
(iv) Do the benefits of incorporating remotely sensedpredictors in invasion risk models differ by species lifeform or by origin status?
MATERIALS AND METHODS
In order to evaluate the potential distributions of selected invasiveplant species in SEA and to assess the contributions of remotelysensed environmental predictors to SDMs, we developed threemodel sets: models constructed along climate data only (CLIM),models with RS only (RS) and models with both climate and RSdata (COMB). CLIM models used well-established bioclimaticdatasets. The compiled RS predictor set covered a diverse rangeof surface parameters, namely topography, soil properties, globalland cover, and vegetation productivity (GPP). Models usedthe Maximum Entropy (MaxEnt) algorithm. Model comparisonswere based on the AUC score of model performance, averagepredicted areas, the level of spatial agreement in predicteddistributions between model results, and the usage of RS andCLIM variables. The evaluation of invasion risk across lifeforms and origin status used predictions of suitable habitatarea for individual species and predicted maps of invader
richness. These datasets and methods are described in more detailbelow.
Study Species and Occurrence DataIn this study, we modeled the potential distributions of 14invasive species (Table 2) identified from the lists of nativeand non-native invasive species known in SEA (Matthews andBrand, 2004) and Vietnam (Ministry of Natural Resourcesand Environment and Ministry of Agriculture and Ruraldevelopment, 2013).
Species occurrences were collected from the GlobalBiodiversity Information Facility1. Records were cleanedfor obvious spatial errors (e.g., points that occurred in the oceanfor terrestrial species) in ArcMap and duplicate records in thedataset were discarded (following Barik and Adhikari, 2011). Allspecies modeled had more than ten occurrence records withinthe study area. The species occurrence records span lengthycollection periods. For each of the 14 species studied, the medianyears of the observations occurred in the period 1956–2005.
Climate DataBioclimatic variables were obtained from the WorldClimdatabase (Version 1.4), interpolated from measurementsrecorded during the period 1960 to 1990 from ∼46,000 climatestations worldwide (Hijmans et al., 2005). Eleven temperatureand eight precipitation metrics, at 1 km resolution, were used,including annual means, seasonality, and extreme or limitingclimatic conditions (Table 3). This dataset has been widely usedfor studies of plant species distributions (Pearson et al., 2007;Hernandez et al., 2008; Cord and Rödder, 2011; Zhu et al., 2017).
Remote Sensing DataA Digital Elevation Model (DEM) was derived from GTOPO302at 30 arc second resolution (approximately 1 km) (USGS,1996). Ten soil layers representing soil physical and chemicalproperties (Hengl et al., 2014) (Table 3) at 1 km resolution wereextracted from ftp://ftp.soilgrids.org/data/archive/12.Apr.2014/.This dataset was empirically developed from global compilationsof publicly available soil profile data (ca. 110,000 soil profiles) anda selection of ∼75 global environmental covariates representingsoil forming factors (mainly MODIS images, climate surfaces,Global Lithological Map, Harmonized World Soil Database andelevation) (Hengl et al., 2014).
We also included the consensus land cover layers developedby Tuanmu and Jetz (2014). They provide a continuous estimateof the probability of the occurrence of each of 12 land coverclasses in each pixel, calculated from the agreements between fourglobal land cover products. These estimates have been shownto have a greater ability to predict species distributions thanthe original categorical land cover products (Tuanmu and Jetz,2014). These land cover data have a 1 km spatial resolution andare available online at http://www.earthenv.org/landcover. Theyrepresent consensus conditions incorporating estimates from the
1http://www.gbif.org/2http://earthexplorer.usgs.gov/
Frontiers in Plant Science | www.frontiersin.org 4 May 2017 | Volume 8 | Article 770
ftp://ftp.soilgrids.org/data/archive/12.Apr.2014/http://www.earthenv.org/landcoverhttp://www.gbif.org/http://earthexplorer.usgs.gov/http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 5
Truong et al. Weed Risk Mapping in Southeast Asia
TAB
LE2
|Des
crip
tio
no
fth
est
udy
spec
ies.
Fam
ilyna
me
Co
mm
on
nam
eS
cien
tifi
cna
me
Life
form
Ori
gin
Med
ian
year
of
ob
serv
atio
ns
Hab
itat
Ast
erac
eae
Sia
mw
eed
Chr
omol
aena
odor
ata
Shr
ubN
on-n
ativ
e20
02•
Hum
idpa
rtof
the
inte
r-tr
opic
alzo
ne,e
leva
tions
belo
w20
00m
•O
pen
seco
ndar
yha
bita
ts
Whi
teto
pW
eed
Par
then
ium
hyst
erop
horu
sH
erb
Non
-nat
ive
2005
•H
umid
and
sub-
hum
idtr
opic
s•
Wid
eva
riety
ofso
ilty
pes,
mor
epr
efer
ably
inhe
avie
rfe
rtile
soils
•D
istu
rbed
habi
tats
(e.g
.,ro
adsi
des,
railw
aytr
acks
,riv
er,a
ndcr
eek
bank
s,bu
ildin
gs)
Mile
-a-M
inut
eM
ikan
iam
icra
ntha
Vin
eN
on-n
ativ
e20
03•
Dam
p,lo
wla
ndcl
earin
gs,o
rop
enar
eas
•S
trea
ms
and
road
side
s,in
orne
arfo
rest
s,fo
rest
plan
tatio
ns,p
astu
res,
fenc
elin
es,t
ree
crop
s
Goa
twee
dA
gera
tum
cony
zoid
esH
erb
Non
-nat
ive
1981
•D
istu
rbed
habi
tats
,roa
dsid
es,d
egra
ded
past
ure
and
culti
vate
dar
eas
Con
volv
ulac
eae
Boi
sM
erre
mia
bois
iana
Vin
eN
ativ
e19
56•
Fore
sts;
elev
atio
nsof
100–
1300
m1
Faba
ceae
Gia
ntse
nsiti
vepl
ant
Mim
osa
dipl
otric
haS
hrub
Non
-nat
ive
2000
•Fe
rtile
area
s;hu
mid
area
sw
ithav
aila
ble
soil
moi
stur
e•
Ope
nan
ddi
stur
bed
habi
tats
Cat
claw
mim
osa
Mim
osa
pigr
aS
hrub
Non
-nat
ive
1999
•R
ipar
ian
area
san
dan
thro
poge
nic
habi
tats
(agr
icul
tura
lare
as)
•D
istu
rbed
and
cons
truc
tion
site
s
Whi
tele
adtr
eeLe
ucae
nale
ucoc
epha
laS
hrub
/Tr
eeN
on-n
ativ
e19
90•
Ope
n,of
ten
coas
talh
abita
ts•
Sem
i-nat
ural
and
dist
urbe
dha
bita
ts
Poa
ceae
Buf
felg
rass
Cen
chru
sec
hina
tus
Gra
ssN
on-n
ativ
e19
70•
Trop
ical
regi
ons
•D
ryan
dm
oist
regi
ons
inra
infe
dar
eas
and
irrig
ated
crop
s•
Mod
erat
em
oist
ure
and
light
,san
dy,w
ell-d
rain
edso
ilsat
low
elev
atio
ns
Bam
boo
gras
sM
icro
steg
ium
cilia
tum
Gra
ssN
ativ
e20
00•
Alo
ngm
esic
road
side
s,ra
ilroa
drig
ht-o
f-w
aydi
tche
s,ut
ility
right
-of-
way
,etc
.Wet
land
,su
cces
sion
alfo
rest
,pla
nted
fore
st,f
ores
tedg
esan
dm
argi
ns,w
oodl
and
bord
ers
•N
otin
area
sw
ithpe
riodi
cst
andi
ngw
ater
,nor
infu
ll,di
rect
sunl
ight
Pol
ygon
acea
eW
ater
hyac
inth
Eich
horn
iacr
assi
pes
Her
bN
on-n
ativ
e19
63•
Trop
ical
and
sub-
trop
ical
fresh
wat
erla
kes
and
river
s,es
peci
ally
thos
een
riche
dw
ithpl
ant
nutr
ient
s,flo
oded
rice
Tam
aric
acea
eLa
ntan
aLa
ntan
aca
mar
aS
hrub
Non
-nat
ive
1982
•D
istu
rbed
area
s,pa
stur
es,r
oads
ides
and
som
etim
esin
nativ
efo
rest
s.
Legu
min
osae
Bau
hini
aB
auhi
nia
tour
anen
sis
Vin
eN
ativ
e19
57•
Ope
nfo
rest
san
dth
icke
tsin
valle
ysan
don
slop
es;5
00–1
200
m2
Kud
zuP
uera
riam
onta
naV
ine
Nat
ive
1983
•W
oods
,pla
ntat
ion
fore
sts,
open
area
s,ab
ando
ned
field
s•
Wid
eva
riety
ofso
ilty
pes
butd
oes
notf
avor
very
wet
soils
•W
ide
geog
raph
ican
dcl
imat
icra
nge
1ht
tp:/
/ww
w.in
atur
alis
t.org
/tax
a/36
3279
-Mer
rem
ia-b
oisi
ana
2ht
tp:/
/ww
w.e
flora
s.or
g/flo
rata
xon.
aspx
?flor
a_id
=3a
ndta
xon_
id=
2000
1196
1O
rigin
and
habi
tat
pref
erre
dar
ecl
assi
fied
acco
rdin
gto
Inva
sive
Spe
cies
Com
pend
ium
deve
lope
dby
CA
BI(
http
://w
ww
.cab
i.org
/isc)
.
Frontiers in Plant Science | www.frontiersin.org 5 May 2017 | Volume 8 | Article 770
http://www.inaturalist.org/taxa/363279-Merremia-boisianahttp://www.efloras.org/florataxon.aspx?flora_id=3andtaxon_id=200011961http://www.cabi.org/ischttp://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 6
Truong et al. Weed Risk Mapping in Southeast Asia
time period 1992–2006, but with greater weight to the later dates(Tuanmu and Jetz, 2014).
To quantify spatial and temporal variation in vegetationproductivity, we used global annual MODIS17A3 (version005) Gross primary productivity (GPP) data for 14 years(2001–2014) at 1 km resolution (Running et al., 2004). ThePrimary Production products are designed to provide an accurateregular measure of the yearly growth of the terrestrial vegetation(Heinsch et al., 2003). Data were downloaded from the NumericalTerradynamic Simulation Group (NTSG) at the Universityof Montana3. The mean and coefficient of variation of GPP(inter-annual variability) were calculated over the time series ateach pixel and supplied to the SDMs.
All predictor variable layers were aligned to a common 1 kmgrid and projected in the Asia South Albers Equal Area Conicsystem using nearest neighbor resampling. Spatial environmentallayers were pre-processed in the TerrSet software (Eastman,2015).
Selection of Environmental PredictionsTo minimize predictor multicollinearity and its impact onsubsequent analyses, we evaluated the inter-correlations amongthe 44 variables for all terrestrial pixels and retained a subsetof uncorrelated (|r| < 0.75) predictor variables for speciesdistribution modeling. Including too much flexibility may makeit difficult for the model to distinguish noise from the true speciesresponse in real data sets (Baldwin, 2009; Merow et al., 2013).Minimizing correlation among variables, therefore, is assumed toincrease the performance of species modeling (Austin, 2002). Inthis way, we reduced the number of predictors used per speciesto 7 climatic (out of 19) and 14 RS (out of 24) variables. All soilestimates were highly correlated across the study area, so only onewas retained. See Table 3 for the full list of initial variables, andthose that were retained for modeling.
Modeling Habitat Suitability of SpeciesTo model habitat suitability, we used MaxEnt (version 3.3.3), ageneral-purpose machine learning method (Phillips et al., 2006).Among species distribution modeling techniques, MaxEnt is oneof the most popular algorithms due to its predictive accuracyand ease of use (Elith et al., 2006; Phillips and Dudík, 2008).There are some characteristics that make MaxEnt highly suitableto modeling species distributions such as use of presence-onlyspecies data, flexibility in the handling of environmental data –including both continuous and categorical variables, and anability to fit complex responses to the environmental variables(Phillips et al., 2006). Notably, MaxEnt is less sensitive to samplesize, which makes MaxEnt a preferred predictive model across allsample sizes (Wisz et al., 2008).
In this study, we developed SDMs based only on theless-correlated climate and/or remotely sensed predictors withMaxEnt. To reduce overfitting, the regularization multiplier wasset at 4. This parameter determines how strongly increases inmodel complexity are penalized during model optimization;higher values produce simpler models that are less overfit to
3http://www.ntsg.umt.edu/project/mod17
TABLE 3 | Environmental variables.
Variables Type of data Source
Bedrock Soil Hengl et al., 2014
Bulk density Soil Hengl et al., 2014
Cation exchange capacity Soil Hengl et al., 2014
Soil texture fraction clay Soil Hengl et al., 2014
Coarse fragments volumetric Soil Hengl et al., 2014
Soil organic carbon stock Soil Hengl et al., 2014
Soil organic carbon content Soil Hengl et al., 2014
Soil pH Soil Hengl et al., 2014
Soil texture fraction silt Soil Hengl et al., 2014
Soil texture fraction sand Soil Hengl et al., 2014
Evergreen/deciduous needleleaf trees
Land cover Tuanmu and Jetz, 2014
Evergreen broadleaf trees Land cover Tuanmu and Jetz, 2014
Deciduous broadleaf trees Land cover Tuanmu and Jetz, 2014
Mixed/other trees Land cover Tuanmu and Jetz, 2014
Shrubs Land cover Tuanmu and Jetz, 2014
Herbaceous vegetation Land cover Tuanmu and Jetz, 2014
Cultivated and managedvegetation
Land cover Tuanmu and Jetz, 2014
Regularly flooded vegetation Land cover Tuanmu and Jetz, 2014Urban/built-up Land cover Tuanmu and Jetz, 2014Snow/ice Land cover Tuanmu and Jetz, 2014
Barren Land cover Tuanmu and Jetz, 2014Open water Land cover Tuanmu and Jetz, 2014Gross primary productivitycoefficient of variation(GPP_CV)
Vegetationproductivity
Heinsch et al., 2003
Gross primary productivity(GPP_Mean)
Vegetationproductivity
Heinsch et al., 2003
Digital elevation model Elevation USGS, 1996
Annual mean temperature Climate Hijmans et al., 2005Mean diurnal temperaturerange
Climate Hijmans et al., 2005
Isothermality Climate Hijmans et al., 2005Temperature seasonality Climate Hijmans et al., 2005
Max temperature of warmestmonth
Climate Hijmans et al., 2005
Min temperature of coldestmonth
Climate Hijmans et al., 2005
Temperature annual range Climate Hijmans et al., 2005
Mean temperature of wettestquarter
Climate Hijmans et al., 2005
Mean temperature of driestquarter
Climate Hijmans et al., 2005
Mean temperature of warmestquarter
Climate Hijmans et al., 2005
Mean temperature of coldestquarter
Climate Hijmans et al., 2005
Annual precipitation Climate Hijmans et al., 2005
Precipitation of wettestmonth
Climate Hijmans et al., 2005
Precipitation of driest month Climate Hijmans et al., 2005
Precipitation seasonality Climate Hijmans et al., 2005
Precipitation of wettest quarter Climate Hijmans et al., 2005
Precipitation of driest quarter Climate Hijmans et al., 2005
Precipitation of warmestquarter
Climate Hijmans et al., 2005
Precipitation of coldest quarter Climate Hijmans et al., 2005
Bold text indicates the variables used as input for MaxEnt modeling.
Frontiers in Plant Science | www.frontiersin.org 6 May 2017 | Volume 8 | Article 770
http://www.ntsg.umt.edu/project/mod17http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 7
Truong et al. Weed Risk Mapping in Southeast Asia
the training data. Radosavljevic and Anderson (2014) foundthat regularization multiplier values from 2.00 to 4.00 weregenerally appropriate to minimize overfitting. For all 14 species,we created 10 random data partitions with 70% of the pointlocalities assigned for training and 30% for testing and ran thethree scenarios (see below) with each of these replicate partitions.Random samples of 10,000 background points were also used todevelop each model.
MaxEnt model performance was evaluated using the areaunder the receiver operating characteristic curve (AUC) assessedon the withheld set of test points. AUC values range from0 to 1. Values of 0.5 indicate that the model performs no betterthan expected by chance, while an AUC of 1 suggests perfectdiscriminatory abilities. Models with AUC > 0.7 are consideredto achieve acceptable performance (Swets, 1988). Mean values,averaging across the 10 replicate runs and across species, of theresulting AUC distributions were used to compare the modelscenarios run with different predictor sets. Continous MaxEntoutputs were converted to binary maps of habitat suitabilityusing the tenth percentile training presence threshold (Escalanteet al., 2013) in order to estimate the area of suitable habitatfor each species predicted by each model. Variable usage by themodels was determined with (1) a variable importance measureestimated as the decrease in model performance when a givenvariable was randomized, and (2) marginal variable responsecurves, which plot the predicted suitability for a species acrossthe range of values for a given variable while all other variablesare held at their mean values.
To test the contribution of RS data to modeling invasivespecies distributions, we ran MaxEnt with climate and satellitelayers in separation and combination. Three scenarios wereevaluated: MaxEnt runs with (1) climate data only (CLIM),these include the three temperature and four precipitation layers
from the final reduced subset; (2) remote sensing data only(RS), with two GPP layers, one soil layer (pH) and eleven landcover classes from the reduced subset; and (3) climate andRS data combined (COMB) using all 21 layers of the reducedsubset (see Table 3). The evaluation was based on (i) the AUCscore; (ii) average predicted areas; (iii) % agreement in predicteddistributions between model results; and (iv) differences invariable importance for the RS and CLIM variables. Thesecomparisons were performed for all species overall, and whengrouping by life forms and origin status. For the assessmentof invasion risk, binary maps of habitat suitability for eachspecies from the COMB model runs were used to determinethe predicted habitat area and combined into maps of invaderrichness to compare the relative level of invasion risk among plantlife forms and native/non-native invasive species.
RESULTS
Model PerformanceOverall, species distributions were generally predictedsuccessfully. All species were successfully modeled (AUC > 0.7)by at least one predictor set (Table 4). Species with few occurrencerecords (less than 20), such as Bauhinia touranensis, Mimosapigra, and Merremia boisiana, tended to be less successfullymodeled in some of the model scenarios (AUC < 0.7). Theremaining species with greater data availability achieved “good”(AUC > 0.8) to “excellent” (AUC > 0.9) performance (Table 4),according to the classification of Swets (1988).
Across all species, the performance of the CLIM and COMBmodels was roughly equivalent (test AUC = 0.84 ± 0.08).Thus, along this metric alone, CLIM models may be preferable,as they are more parsimonious. On average, the RS models
TABLE 4 | Variability (mean and standard devation) of species-specific AUC (area under the curve) scores, evaluated against the withheld test set of 30%of the presence records, for fourteen invasive weeds in 10 partition runs.
Species Number ofoccurrences
CLIM RS COMB
Ageratum conyzoides 360 0.81 ± 0.01 0.74 ± 0.02 0.84 ± 0.02
Bauhinia touranensis 19 0.85 ± 0.03 0.51 ± 0.16 0.76 ± 0.07
Cenchrus echinatus 110 0.85 ± 0.04 0.88 ± 0.03 0.86 ± 0.04
Chromolaena odorata 167 0.88 ± 0.03 0.77 ± 0.03 0.89 ± 0.03
Eichhornia crassipes 81 0.65 ± 0.05 0.84 ± 0.04 0.84 ± 0.06
Lantana camara 162 0.90 ± 0.02 0.77 ± 0.04 0.88 ± 0.02
Leucaena leucocephala 192 0.85 ± 0.03 0.82 ± 0.02 0.87 ± 0.03
Merremia boisiana 13 0.74 ± 0.08 0.50 ± 0.10 0.72 ± 0.07
Microstegium ciliatum 96 0.86 ± 0.03 0.72 ± 0.06 0.86 ± 0.03
Mikania micrantha 171 0.92 ± 0.02 0.81 ± 0.04 0.93 ± 0.02
Mimosa diplotricha 54 0.86 ± 0.06 0.78 ± 0.07 0.85 ± 0.05
Mimosa pigra 19 0.73 ± 0.09 0.66 ± 0.06 0.64 ± 0.09
Parthenium hysterophorus 76 0.97 ± 0.02 0.85 ± 0.04 0.97 ± 0.01
Pueraria montana 417 0.89 ± 0.02 0.83 ± 0.02 0.84 ± 0.03
Mean 0.84 ± 0.08 0.75 ± 0.12 0.84 ± 0.08
Three variable sets were used for each species. CLIM includes only bioclimatic predictors; RS includes only remote sensing predictors; COMB includes variables in CLIMand RS. AUC values for the best-performing model for each species are indicated in bold.
Frontiers in Plant Science | www.frontiersin.org 7 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 8
Truong et al. Weed Risk Mapping in Southeast Asia
FIGURE 1 | Test AUC by life forms (A) and by origin (B) among models. CLIM includes only bioclimatic predictors; RS includes only remote-sensing predictors;COMB includes variables in CLIM and RS. The error bars are standard deviations.
were the least successful (test AUC = 0.75 ± 0.12) (Table 4).However, the rankings differed somewhat for individual speciesand between species categories. CLIM models were preferredfor 8 species, RS for 2, and COMB for the remaining4 (Table 4). RS models were found to perform worst inpredicting vine species (Figure 1) and native invasive species(Figure 1).
COMB models generally predicted smaller areas of suitablehabitat than either CLIM or RS models. This pattern wasconsistent across life forms and origin status, but strongest forherbs, shrubs, and non-native invasive species (Figure 2). CLIMand RS models tended to predict similar areas of suitable habitat,except for the case of vines and native invasive species. The RSmodels for these groups predicted larger areas of suitable habitatthan did CLIM models (Figure 2).
In general, spatial agreement in predicted habitat was greatestfor pairwise comparisons with the COMB models (Figure 3).As an exception to this pattern, the agreement between COMBand RS was as low as between CLIM and RS for vines andnative invasive species. At the individual species level (SupportingInformation S2), COMB tended to be most similar to theindividual model set (CLIM or RS) that performed better in theAUC evaluations (Table 4) – typically CLIM.
The average relative variable importance variedconsiderably among the predictors within the variablesets. In the CLIM set, mean diurnal temperature range(importance = 32.5% ± 22.0 and precipitation of warmestquarter (importance = 23.8% ± 17.4) were most important(Table 5). On average, other temperature variables (isothermalityand annual mean temperature) have an importance around12–13% and other variables contributed less than 10%. Of thevariables in the RS predictor set, herbaceous vegetation landcover (importance = 16.7% ± 8.8) was the most important.Evergreen broadleaf tree, cultivated vegetation and GPP_CVwere also important variables, with permutation importanceranging from 10 to 12% on average. In the COMB predictorset, the contribution of variables was similar to the CLIM andRS scenarios (Table 5). All variables had reduced importancein COMB than in either CLIM or RS, due to the inclusion of alarger number of variables in these models, but the rankings ofvariables within each predictor were generally consistent.
Habitat SuitabilityTo assess the habitat suitability of species, we used resultsfrom COMB models. Response curves of each species (responsecurves are provided in Figure 4 for a selected species ofeach life form that was best modeled by the COMB variableset, and for all species in Supporting Information S1) inCOMB models reveal that, across species, sites were generallypredicted to have high suitability (>0.6) in areas with low meandiurnal temperature range and moderate to high isothermality.The highest suitability (0.9–1) was also generally found inareas with high precipitation in the warmest season. Manymodeled species (Chromolaena odorata, Cenchrus echinatus,Eichhornia crassipes, Lantana camara, Mimosa diplotricha) werenot predicted to invade closed areas such as forests (negativeresponses to high canopy land-cover classes), although theaggressive vine Pueraria montana is a notable exception. Inaddition, for species models with important contributions fromthe productivity variables, suitability was generally found to behighest in environments with high GPP and low variability ofGPP (Supporting Information S1).
Herb species receive the greatest area predicted to be atrisk of invasion by one or more species (5.3 million km2,versus 4.9 million km2 and 4.3 million km2 for shrubsand vines, respectively), however, the area vulnerable to thegreatest invader richness is fairly concentrated around thenorth and north center of Vietnam (Figure 5). Responsecurves of herb species (Ageratum conyzoides, Cenchrus echinatus,Microstegium ciliatum, and Parthenium hysterophorus) indicatethey prefer high rainfall in the warmest quarter (more than>1500 mm), however, this variable was generally less importantfor herbs than it was for other life forms (SupportingInformation S1). Additionally, herb species prefer habitat withdiurnal temperature ranges less than 10◦C and isothermalityfrom 20 to 70%. Of the land cover variables, invasibility toherbs was more strongly related to the evergreen broadleaf andmixed forest classes, and to the cultivated class than were theother life forms. Response curves indicated that relationshipswith these cover classes were generally negative (SupportingInformation S1).
Shrub species were predicted to have the greatest area atrisk from multiple invaders: 1.3 million km2 were predicted
Frontiers in Plant Science | www.frontiersin.org 8 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 9
Truong et al. Weed Risk Mapping in Southeast Asia
FIGURE 2 | Average predicted area by life forms (A) and by origin (B) among models. Predicted value is identified based on 10% logistic threshold. CLIMincludes only bioclimatic predictors; RS includes only remote-sensing predictors; COMB includes variables in CLIM and RS. The error bars are standard deviations.
FIGURE 3 | Percentage of agreement in predicted area by life forms (A) and by origin (B) among models. CLIM includes only bioclimatic predictors; RSincludes only remote-sensing predictors; COMB includes variables in CLIM and RS. The error bars are standard deviations.
to be suitable for four or more shrub species, as opposedto only 0.6 million km2 for herbs and 86 thousand km2for vines (although note that only four vine species weremodeled). Unlike the other life forms, regions suitable formultiple shrub invaders extended into countries in the southof the region such as Indonesia, Malaysia, and Philippines, aswell as west to Bangladesh (Figure 5). Diurnal temperaturerange and precipitation of the warmest quarter were themost important factors for the distribution of these shrubspecies (e.g., Chromolaena odorata, Lantana camara, Leucaenaleucocephala). Overall, models were more influenced by RSvariables, especially land cover, for shrub species than for theother life forms. Shrubs exhibited generally negative associationswith forested habitat (for all classes except the mixed forests) aswell as with herbaceous land cover (Supporting Information S1).
In contrast to the other groups, large areas were predictedto be invasible to a single vine species. Areas vulnerable togreater richness of invasive vines were much more restricted,tending to occur in north and north-central Vietnam and Taiwan(Figure 5). While Mikania micrantha and Pueraria montanahave less predicted area in SEA, Bauhinia touranensis andMerremia boisiana were predicted to invade much of the region(Supporting Information S2), especially in south China and north
Vietnam. Unlike herbs and shrubs, distributions of vine specieswere generally unrelated to land cover (except for moderateinfluences of herbaceous land cover). Vine species receivedgreater importance of climate factors, especially variables relatedto precipitation, than did the other life forms (SupportingInformation S1).
Results of average predicted area at the species level showedthat as large areas are vulnerable to invasion by native asnon-native invasive species (ca. 2 million km2) over the wholeregion (Figure 2). Cumulative levels of invasion risk are difficultto compare, since over twice as many non-native than nativespecies were modeled, but substantial areas are at risk ofinvasion by one or more species of each origin status (6 millionkm2 and 4.3 million km2, for non-native and native invasivespecies, respectively). Native invasive species richness was mainlyconcentrated in the north and north center of Vietnam; non-native species had wider range of distribution and may potentiallyinvade the whole region (Figure 6).
Comparing the total area predicted by the COMB models tobe susceptible to the invasion of the 14 invasive species suggestswhich of the modeled species may be the greatest threats to theregion. Ageratum conyzoides, Eichhornia crassipes, Leucaeanaleucocephala and Microstegium ciliatum had the highest
Frontiers in Plant Science | www.frontiersin.org 9 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 10
Truong et al. Weed Risk Mapping in Southeast Asia
TABLE 5 | Summary of the mean permutation importance (PI) of fourteeninvasive plant species.
COMB CLIM RS
Mean ± SD Mean ± SD Mean ± SD
GPP_CV 2.1 ± 2.49 10.76 ± 10.24
GPP_Mean 2.83 ± 3.21 8.41 ± 8.2
Soil pH 1.32 ± 0.95 2.51 ± 5.34
Barren 1.21 ± 1.2 2.63 ± 2.09
Cultivated vegetation 3.83 ± 5.64 11.22 ± 7.51
Deciduous broad leaf trees 5.17 ± 4.81 8.86 ± 8.25
Evergreen broad leaf trees 7.1 ± 9.11 12.37 ± 9.93
Evergreen needle leaf trees 4.42 ± 9.24 6.19 ± 9.4
Herbaceous vegetation 7.05 ± 7.38 16.71 ± 8.62
Mixed trees 3.7 ± 4.99 8.46 ± 6.18
Open water 0.79 ±0.8 1.2 ± 0.77
Regular flooded vegetation 0.98 ± 1.6 2.53 ± 4.86
Shrubs 1.77 ± 1.46 6.56 ± 9.19
Urban 1.07 ± 1.19 1.6 ± 1.49
Annual mean temperature 4.32 ± 6.57 13.27 ± 14.57
Mean diurnal temperature range 17.65±16.04 32.48 ± 22.02
Isothermality 7.72 ±6.84 12.46 ± 10.98
Annual precipitation 7.53 ± 14.12 9.06 ± 13.86
Precipitation of wettest month 1.54 ± 1.94 3.26 ± 2.52
Precipitation seasonality 3.67 ± 4.9 5.66 ± 6.56
Precipitation of warmest quarter 14.23 ± 9.93 23.81 ± 17.41
SD is standard deviation. Mean values were calculated from the average of 14species. Values in bold indicate variables with above-average importance in COMB(4.8%), CLIM (14.3%), and RS (7.1%).
predicted area. Lantana camara and Mimosa diplotricha followednext. Parthenium hysterophorus had the lowest predicted area(Supporting Information S2).
DISCUSSION
Model PerformanceQuantitative comparisons of models with various predictorsets showed that models built with incorporation of RS andclimatic data layers substantially reduced predicted areas acrossall life forms and origin status compared to models withclimate and RS data alone (Figure 2). The mapped predictionsfor individual species reflect this pattern spatially (SupportingInformation S2). Suitable habitat modeled with climate variablesalone are quite smooth and generalized, while the inclusionof remotely sensed predictor variables adds more nuancedspatial detail to this overall pattern. The most widely usedbioclimatic predictors, including those evaluated in this study, arederived from station data; interpolation introduces smoothing,producing generalized portrayals of environmental variability.As well, climate generally varies continuously over broad spatialscales. Thus, exclusively climate-based distribution models areunable to capture variations of species diversity at the landscapelevel (Saatchi et al., 2008). As a consequence, large areas ofpredicted suitability are often seen (Thuiller et al., 2004). Incontrast, while the biotic niche axes estimated by RS can further
inform distribution models and enable dynamic models, they areunable to replace climatic factors in identifying suitable habitatas bioclimatic conditions are still essential driving factors forspecies distributions (Thuiller et al., 2004; Cord and Rödder,2011). The high percentage agreement of spatial predictionsbetween models based on climatic predictors only and climaticand RS predictors found in this study, as well as the high variableimportance scores given to climatic predictors in the combinedmodels, also supports the indispensability of climate in shapingthe distribution of invasive plant species. Similar studies have alsofound that using either climatic-derived or RS-derived predictorsalone often leads to the overprediction of species distributions(Buermann et al., 2008; Saatchi et al., 2008; Cord and Rödder,2011; Cord et al., 2014a). By incorporating complementarylimiting environmental conditions, combined models of climaticand remotely sensed predictor variables reduce predicted areas,thereby refining modeled species distributions.
Although clearly refining the spatial patterns of predictedspecies distributions, in general, COMB models did not achievehigher accuracy than models with climate variables alone;RS models were often relatively poor. These results are inline with other studies (Zimmermann et al., 2007; Cord andRödder, 2011; Cord et al., 2014a) that found that models basedon RS data had the lowest AUC, compared to models withclimate-derived predictors and climatic and RS predictors. Someexplanations can be proposed for this. First, there may betemporal mismatch between occurrence data and environmentaldata. This is likely to be a more severe problem for remotelysensed predictors, which generally capture snapshots in time,rather than climatological averages, and which often describeenvironmental conditions, such as vegetation patterns, thatvary over shorter time frames than does climate. Many of theoccurrence records within museum or herbarium collections,comprising GBIF, are older; the land cover and vegetationproductivity present at those sites at the time of the species’presence may not be represented by remotely sensed currentconditions. To test for this problem, we repeated our modelswith recent records only (collected after 1992). Removing olderspecies records reduced model performance overall, likely dueto the much smaller samples available to train the models.Remotely sensed predictors received slightly higher importancevalues in the COMB models than previously, but were stillsecondary to climatic variables (Supporting Information S3).Although temporal correspondence among species occurrencesand environmental variables is a concern and should beconsidered in further studies, it does not seem to contribute toour conclusions.
Alternatively, the quality and information content of the RSproducts may influence model performance. The consensus landcover product was used in this study because it was expectedto be more reliable than traditional global land cover datasets.Additionally, its continuous estimates of the probability of classpresence may avoid errors associated with categorical data andprovide some level of subpixel land cover information. However,it still has limitations related to the input datasets. Global landcover products are constrained to a relatively simple legend, withbroad classes. The consensus product is further constrained to
Frontiers in Plant Science | www.frontiersin.org 10 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 11
Truong et al. Weed Risk Mapping in Southeast Asia
FIGURE 4 | Marginal response curves of Ageratum conyzoides (a non-native herb best modeled by COMB), Leucaena leucocephala (a non-nativeshrub best modeled by COMB) and Mikania micrantha (a non-native vine best modeled by COMB) for variables with importance >5% for eachspecies in COMB models. The orange curve in each plot is average response curve and the blue is standard deviation across all 10 partition runs. See otherspecies in Supporting Information S1.
Frontiers in Plant Science | www.frontiersin.org 11 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 12
Truong et al. Weed Risk Mapping in Southeast Asia
FIGURE 5 | Maps of predicted richness of invasive species by life form produced with COMB set (combing climate and remote sensing data).(A) Herb, (B) Shrub and (C) Vine. The browner the color, the higher the predicted richness of invasive species.
a simplified legend that harmonizes each of the input products.The generality of these classes may not capture regionally relevantdifferences and limit their usefulness to SDMs. The consensusland cover product is also limited by quality of the individual
products it integrates (Tuanmu and Jetz, 2014). In land coverproducts, classification errors are not evenly distributed acrossspace and classes (Strahler et al., 2006). For instance, loweraccuracy for land cover classes of GlobCover products was found
Frontiers in Plant Science | www.frontiersin.org 12 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 13
Truong et al. Weed Risk Mapping in Southeast Asia
FIGURE 6 | Maps of predicted richness of invasive species by origin produced with COMB set (combing climate and remote sensing data). (A) Native,(B) Non-native species. The browner color, the higher predicted richness of invasive species.
FIGURE 7 | Uncertainty in global land cover products revealed by the maximum class probability value, excluding the open water class, received in apixel in the Consensus Land Cover dataset (Tuanmu and Jetz, 2014). Low maximum probability values indicate a great deal of disagreement betweenindividual land cover products.
Frontiers in Plant Science | www.frontiersin.org 13 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 14
Truong et al. Weed Risk Mapping in Southeast Asia
in some areas with limited data coverage (e.g., some areas inAmazonia) or in rugged terrain such as Laos (Bicheron et al.,2008). Also, cloud cover reduces the quality of the RS data,especially in tropical regions (Bradley and Fleishman, 2008).
Classification errors do seem to be contributing to theperformance of RS variables in our study. Unexpectedly, speciesassociations with land cover classes, when they were found tobe important to models, were overwhelmingly negative. Thereis no ecological or logical reason for this. Instead, because theconsensus land cover product estimates the certainty that a classis present, given the individual land cover datasets, this suggeststhat habitat suitability tends to be greatest for the modeled speciesin areas with high land cover uncertainty. Such uncertaintymay be due to inadequacies in the class definitions in thisregion, fine-scaled mosaics of land cover classes within a 1 kmpixel, or simply poor classification performance. Indeed, usingthe maximum estimated probability of class membership as anindicator of certainty supports this interpretation. Large areas ofSEA, including many of the same locations with high-predictedinvasibility, exhibit low certainty of the land cover information(Figure 7). Further work is necessary to validate the consensusland cover products in SEA and, especially, to determine themeaning of areas with great class uncertainty. This is troublingand argues against the use of global land cover products inSDMs. Quantitative remotely sensed estimates of ecosystemstructure and function may overcome some of the problems ofcategorical datasets, and we strongly advocate for their expandeduse and continued evaluation in SDM contexts. Interestingly,the quantitative measures of vegetation productivity used in thisstudy, while making important contributions to the RS modelset, generally dropped out of the COMB models. This may bebecause of interdependencies between climate variables and thephotosynthetic efficiency term used in the MODIS GPP product,which relies on both temperature and moisture (Running et al.,2004), and thus would not be detected by the simple univariatecorrelation analysis used to screen input variables.
Another limitation to model performance in this study isthe sample size of the species occurrence records. Performanceof SDMs in the study varied among species. Species with fewoccurrence records occurring in a wide range of habitats, suchas Mimosa pigra, have lower performance than others. Thisis because SDMs perform better with larger sample sizes andfor species occupying a narrow environmental niche than forgeneralist species (Hernandez et al., 2006). Although Mimosapigra has been recorded as one of the most invasive plants inmany countries in SEA (Thi, 2000; MacKinnon, 2002; Vannaand Nang, 2005; Nghiem et al., 2013), the number of occurrencerecords of this species in SEA is still limited. This reflectslack of research and awareness of the public and governmentfor invasive species detection in the region, which shouldbe more encouraged. Also, using hyperspectral RS to detectinvasive species occurrences (Andrew and Ustin, 2008; Hestiret al., 2008) can be a solution for developing high-quality,unbiased occurrence data inputs (He et al., 2015), and alsomay reduce temporal mismatch between species occurrencesand environmental variables. In addition to model development,sample size influences model evaluation. Performance measures
such as the AUC provide a single spatial summary value. AUC hasbeen criticized for its inability to convey information about thespatial pattern of predictions or uncertainty (Franklin, 2010a).Yet spatial variation can be considerable. Because AUC is oftencalculated from a tiny proportion of the pixels modeled, wildlydifferent spatial predictions can receive similar, and indeed veryhigh, AUC estimates (Synes and Osborne, 2011). For this reason,we prefer to present a suite of evaluation tools, including totalpredicted area and estimates of spatial agreement, in addition tothe AUC.
Habitat SuitabilityBoth non-native and native invasive species were predicted tooccur across large areas of SEA, and thus may pose similarrisk to the region. Among life forms, shrub species potentiallypose greater risk because of the predictions of high shrubinvader richness over large areas, based on the set of speciesassessed. Most countries in the region have suitable habitat forthese species. In general, shrubs exhibited weaker environmentalassociations than the other life forms (as seen in the lowervariable importance scores), suggesting they may be tolerantof a broader range of conditions. Relative to shrub and herbspecies, vine species’ distributions were most strongly driven byclimatic factors. This may facilitate their spread under climatechange. Invasive species may disproportionately benefit fromglobal climate change (Dukes and Mooney, 1999), and vinesmay be a good example of these concerns. Climate projectionsfor the region include increases in annual temperature and insummertime precipitation (Christensen et al., 2007), the lattervariable was important to nearly all vine species distributions, allof which showed positive associations. Without strong controlsby biotic factors such as land cover, vines may invade valuableevergreen broadleaf trees forests in SEA. A native vine, Merremiaboisiana is an example. In the past decade, the vine has spreaddramatically over South China (Wang et al., 2005; Wu et al.,2007) and the north and center of Vietnam (Le et al., 2012) andour results reveal that more than 1.6 million km2 are invasibleto this species, largely concentrated in China and Vietnam.These findings suggest that awareness of invasive species andprevention and eradication efforts should not overlook the lifeform or origin status of the species of concern.
Interestingly, in contrast to our expectations, we found thatfor some species (Microstegium ciliatum and Mimosa diplotricha)suitability was negatively related to the variability of GPP(GPP_CV), which was used to proxy disturbance processes. Thissuggests that invasion is possible even with low disturbance,contradicting knowledge summarized by Lozon and MacIsaac(1997) that the establishment and spread of invasive plants areassociated with disturbance. Although disturbance is certainly afactor in many invasions, an over-generalization that invasionrequires disturbance can lead to low awareness of invasion inintact areas. Further field-based studies about invasibility of thesespecies under difference disturbance levels should be conducted.The effectiveness of GPP variability as an indicator of diversedisturbance processes and diverse ecosystems should also beevaluated. The relatively short duration of the satellite archivefrom which it was computed is certainly a limitation.
Frontiers in Plant Science | www.frontiersin.org 14 May 2017 | Volume 8 | Article 770
http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 15
Truong et al. Weed Risk Mapping in Southeast Asia
Given that many of the study species were identified fromVietnam’s invasive weed list, it is not surprising that we found,within the region, north and north central Vietnam were mostsusceptible to the invasion of weeds (Figures 5, 6). However, it isworth emphasizing that many of the invasive weeds predicted inthis region also have high invasibility in China, where outbreakshave been recorded (Yan et al., 2001). Biological invasions are atrans-border issue. Similarly, provinces (Guangxi, Quangdong,and Yunnan) sharing borders with Vietnam, Lao, and Myanmarare listed as areas with a high number of invasive species in China(Xu et al., 2012). Effective management requires that invasionsbe considered in the context of the region (SEA), rather than acountry (Paini et al., 2010). Studies such as ours can help theVietnamese and other governments to prioritize managementactions for invasive species within the country and also to informbiosecurity policy across borders.
CONCLUSION
This study demonstrated that although the environmentalattributes derived from RS data did not strongly improve theaccuracy of SDM predictions, they did provide more landscape-level detail that refined species distribution predictions inspace. Therefore, the inclusion of remotely sensed variables inSDMs likely is worthwhile. Furthermore, our results highlightshortcomings of land cover products, which are widely usedin SDMs. There are widespread uncertainties in global landcover products and, disconcertingly, those sites with the greatestuncertainty also seem to be consistently ecologically importantto the modeled species. We caution against continued use of
land cover information in SDMs, which may propagate errorsand confound interpretation. Greater adoption of quantitativeremotely sensed datasets estimating ecosystem structure andfunction may mitigate the weaknesses and limited utility ofRS observed in this study. From the standpoint of biodiversitymanagement, our findings have implications in targetingmanagement to susceptible areas, providing initial data forinvasive species risk assessments, and proposing biosecuritypolicy in the region.
AUTHOR CONTRIBUTIONS
TT prepared input data, performed models and interpretedresults, wrote manuscript and acted as corresponding author. MAsupervised development of work, provided guidance throughoutthe project, and edited manuscript. GH contributed to editingmanuscript.
FUNDING
TT was supported by Australia Awards Scholarship for her Ph.D.studies.
SUPPLEMENTARY MATERIAL
The Supplementary Material for this article can be found onlineat: http://journal.frontiersin.org/article/10.3389/fpls.2017.00770/full#supplementary-material
REFERENCESAndrew, M. E., and Ustin, S. L. (2008). The role of environmental context in
mapping invasive plants with hyperspectral image data. Remote Sens. Environ.112, 4301–4317. doi: 10.1016/j.rse.2008.07.016
Andrew, M. E., and Ustin, S. L. (2009). Habitat suitability modelling of aninvasive plant with advanced remote sensing data. Divers. Distrib. 15, 627–640.doi: 10.1111/j.1472-4642.2009.00568.x
Andrew, M. E., Wulder, M. A., and Nelson, T. A. (2014). Potential contributionsof remote sensing to ecosystem service assessments. Progr. Phys. Geogr. 38,328–353. doi: 10.1177/0309133314528942
Austin, M. P. (2002). Spatial prediction of species distribution: an interfacebetween ecological theory and statistical modelling. Ecol. Model. 157, 101–118.doi: 10.1016/S0304-3800(02)00205-3
Avril, L., and Kelty, M. J. (1999). Establishment and control of hay-scentedfern: a native invasive species. Biol. Invasions 1, 223–236. doi: 10.1023/A:1010098316832
Baldwin, R. A. (2009). Use of maximum entropy modeling in wildlife research.Entropy 11, 854–866. doi: 10.3390/e11040854
Barik, S., and Adhikari, D. (2011). “Predicting geographic distribution of aninvasive species (Chromolaena odorata L. (King) & H. E. Robins) in the Indiansubcontinent under climate change scenarios,” in Invasive Alien Plants—AnEcological Appraisal for the Indian Sub-continent, eds J. R. Bhatt, J. S. Singh,S. P. Singh, R. S. Tripathi, and R. K. Kohli (Wallingford: CABI), 77–88. doi:10.1079/9781845939076.0077
Bear, R., Hill, W., and Pickering, C. M. (2006). Distribution and diversity ofexotic plant species in montane to alpine areas of Kosciuszko National Park.Cunninghamia 9, 559–570.
Bicheron, P., Defourny, P., Brockmann, C., Schouten, L. I., Vancutsem, C.,Huc, M., et al. (2008). GLOBCOVER: Products Description and ValidationReport. Toulouse: MEDIAS France.
Bradley, B. A., and Fleishman, E. (2008). Can remote sensing of land cover improvespecies distribution modelling? J. Biogeogr. 35, 1158–1159. doi: 10.1111/j.1365-2699.2008.01928.x
Bradley, B. A., and Mustard, J. F. (2006). Characterizing the landscape dynamicsof an invasive plant and risk of invasion using remote sensing. Ecol. Appl. 16,1132–1147. doi: 10.1890/1051-0761(2006)016[1132:CTLDOA]2.0.CO;2
Bradley, B. A., Olsson, A. D., Wang, O., Dickson, B. G., Pelech, L., Sesnie,S. E., et al. (2012). Species detection vs. habitat suitability: are we biasinghabitat suitability models with remotely sensed data? Ecol. Model. 244, 57–64.doi: 10.1016/j.ecolmodel.2012.06.019
Buermann, W., Saatchi, S., Smith, T. B., Zutta, B. R., Chaves, J. A., Milá, B., et al.(2008). Predicting species distributions across the Amazonian and Andeanregions using remote sensing data. J. Biogeogr. 35, 1160–1176. doi: 10.1111/j.1365-2699.2007.01858.x
Burke, M. J. W. (1996). An experimental study of plant community invasibility.Ecology 77, 776–790. doi: 10.2307/2265501
Campos, V. E., Cappa, F. M., Viviana, F. M., and Giannoni, S. M. (2016). Usingremotely sensed data to model suitable habitats for tree species in a desertenvironment. J. Veg. Sci. 27, 200–210. doi: 10.1111/jvs.12328
Christensen, J. H., Hewitson, B., Busuioc, A., Chen, A., Gao, X., Held, I., et al.(2007). “Regional climate projections,” in Climate Change 2007: The PhysicalScience Basis. Contribution of Working Group I to the Fourth Assessment Reportof the Intergovernmental Panel on Climate Change, eds S. Solomon, D. Qin, M.Manning, Z. Chen, M. Marquis, K. B. Averyt, et al. (Cambridge: CambridgeUniversity Press).
Frontiers in Plant Science | www.frontiersin.org 15 May 2017 | Volume 8 | Article 770
http://journal.frontiersin.org/article/10.3389/fpls.2017.00770/full#supplementary-materialhttp://journal.frontiersin.org/article/10.3389/fpls.2017.00770/full#supplementary-materialhttps://doi.org/10.1016/j.rse.2008.07.016https://doi.org/10.1111/j.1472-4642.2009.00568.xhttps://doi.org/10.1177/0309133314528942https://doi.org/10.1016/S0304-3800(02)00205-3https://doi.org/10.1023/A:1010098316832https://doi.org/10.1023/A:1010098316832https://doi.org/10.3390/e11040854https://doi.org/10.1079/9781845939076.0077https://doi.org/10.1079/9781845939076.0077https://doi.org/10.1111/j.1365-2699.2008.01928.xhttps://doi.org/10.1111/j.1365-2699.2008.01928.xhttps://doi.org/10.1890/1051-0761(2006)016[1132:CTLDOA]2.0.CO;2https://doi.org/10.1016/j.ecolmodel.2012.06.019https://doi.org/10.1111/j.1365-2699.2007.01858.xhttps://doi.org/10.1111/j.1365-2699.2007.01858.xhttps://doi.org/10.2307/2265501https://doi.org/10.1111/jvs.12328http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 16
Truong et al. Weed Risk Mapping in Southeast Asia
Cord, A., and Rödder, D. (2011). Inclusion of habitat availability in speciesdistribution models through multi-temporal remote-sensing data? Ecol. Appl.21, 3285–3298. doi: 10.1890/11-0114.1
Cord, A. F., Klein, D., Gernandt, D. S., la Rosa, J. A. P., and Dech, S. (2014a).Remote sensing data can improve predictions of species richness by stackedspecies distribution models: a case study for Mexican pines. J. Biogeogr. 41,736–748. doi: 10.1111/jbi.12225
Cord, A. F., Klein, D., Mora, F., and Dech, S. (2014b). Comparing the suitabilityof classified land cover data and remote sensing variables for modelingdistribution patterns of plants. Ecol. Model. 272, 129–140. doi: 10.1016/j.ecolmodel.2013.09.011
D’Antonio, C. M. (1993). Mechanisms controlling invasion of coastal plantcommunities by the alien succulent Carpobrotus edulis. Ecology 74, 83–95.doi: 10.2307/1939503
Davis, M. A., Grime, J. P., and Thompson, K. (2000). Fluctuating resourcesin plant communities: a general theory of invasibility. J. Ecol. 88, 528–534.doi: 10.1046/j.1365-2745.2000.00473.x
Deblauwe, V., Droissart, V., Bose, R., Sonké, B., Blach-Overgaard, A., Svenning,J. C., et al. (2016). Remotely sensed temperature and precipitation data improvespecies distribution modelling in the tropics. Glob. Ecol. Biogeogr. 25, 443–454.doi: 10.1111/geb.12426
Dormann, C. F., Schymanski, S. J., Cabral, J., Chuine, I., Graham, C., Hartig, F.,et al. (2012). Correlation and process in species distribution models: bridging adichotomy. J. Biogeogr. 39, 2119–2131. doi: 10.1111/j.1365-2699.2011.02659.x
Dukes, J. S., and Mooney, H. A. (1999). Does global change increase the success ofbiological invaders? Trends Ecol. Evol. 14, 135–139.
Eastman, J. (2015). TerrSet: Geospatial Monitoring and Modeling Software.Worcester, MA: Clark University.
Elith, J., Graham, C. H., Anderson, R. P., Dudík, M., Ferrier, S., Guisan, A.,et al. (2006). Novel methods improve prediction of species’ distributionsfrom occurrence data. Ecography 29, 129–151. doi: 10.1111/j.2006.0906-7590.04596.x
Engler, R., Waser, L. T., Zimmermann, N. E., Schaub, M., Berdos, S., Ginzler, C.,et al. (2013). Combining ensemble modeling and remote sensing for mappingindividual tree species at high spatial resolution. For. Ecol. Manag. 310, 64–73.doi: 10.1016/j.foreco.2013.07.059
Escalante, T., Rodríguez-Tapia, G., Linaje, M., Illoldi-Rangel, P., and González-López, R. (2013). Identification of areas of endemism from species distributionmodels: threshold selection and Nearctic mammals. TIP 16, 5–17. doi: 10.1016/S1405-888X(13)72073-4
Estes, L., Okin, G., Mwangi, A., and Shugart, H. (2008). Habitat selection by a rareforest antelope: a multi-scale approach combining field data and imagery fromthree sensors. Remote Sens. Environ. 112, 2033–2050. doi: 10.1016/j.rse.2008.01.004
Evangelista, P., Stohlgren, T., Morisette, J., and Kumar, S. (2009). Mapping invasivetamarisk (Tamarix): a comparison of single-scene and time-series analyses ofremotely sensed data. Remote Sens. 1, 519–533. doi: 10.3390/rs1030519
Feilhauer, H., He, K. S., and Rocchini, D. (2012). Modeling species distributionusing niche-based proxies derived from composite bioclimatic variables andMODIS NDVI. Remote Sens. 4:2057. doi: 10.3390/rs4072057
Fernández, M., Hamilton, H., Alvarez, O., and Guo, Q. (2012). Does adding multi-scale climatic variability improve our capacity to explain niche transferability ininvasive species? Ecol. Model. 246, 60–67. doi: 10.1016/j.ecolmodel.2012.07.025
Fox, M. D., and Fox, B. J. (1986). “The susceptibility of natural communities toinvasion,” in Ecology of Biological Invasions, eds R. H. Groves and J. J. Burdon(Cambridge: Cambridge University Press), 57–66.
Franklin, J. (2010a). Mapping Species Distributions: Spatial Inference and Prediction.Cambridge: Cambridge University Press. doi: 10.1017/CBO9780511810602
Franklin, J. (2010b). Moving beyond static species distribution models in supportof conservation biogeography. Divers. Distrib. 16, 321–330. doi: 10.1111/j.1472-4642.2010.00641.x
Garrard, G., Bekessy, S., and Wintle, B. (2009). Determining Necessary SurveyEffort to Detect Invasive Weeds in Native Vegetation Communities. Final ReportACERA Project No. 0906. Parkville, VIC: University of Melbourne.
Genovesi, P. (2005). Eradications of invasive alien species in Europe: a review. Biol.Invasions 7, 127–133. doi: 10.1007/s10530-004-9642-9
Goetz, S. J., Bond-Lamberty, B., Law, B. E., Hicke, J. A., Huang, C., Houghton, R. A.,et al. (2012). Observations and assessment of forest carbon dynamics following
disturbance in North America. J. Geophys. Res. Biogeosci. 117, G02022.doi: 10.1029/2011jg001733
Gonçalves, J., Alves, P., Pôças, I., Marcos, B., Sousa-Silva, R., Lomba, Â., et al.(2016). Exploring the spatiotemporal dynamics of habitat suitability to improveconservation management of a vulnerable plant species. Biodiver. Conserv. 25,2867–2888. doi: 10.1007/s10531-016-1206-7
Gower, D., Johnson, K., Richardson, J., Rosen, B., Rüber, L., and Williams, S. (2012).Biotic Evolution and Environmental Change in Southeast Asia. Cambridge:Cambridge University Press. doi: 10.1017/CBO9780511735882
Groves, R., Hosking, J., Batianoff, G., Cooke, D., Cowie, I., Johnson, R., et al.(2003). Weed Categories for Natural and Agricultural Ecosystem Management.Canberra, ACT: Bureau of Rural Sciences.
Guisan, A., and Thuiller, W. (2005). Predicting species distribution: offering morethan simple habitat models. Ecol. Lett. 8, 993–1009. doi: 10.1111/j.1461-0248.2005.00792.x
Guisan, A., and Zimmermann, N. E. (2000). Predictive habitat distribution modelsin ecology. Ecol. Model. 135, 147–186. doi: 10.1016/S0304-3800(00)00354-9
Harrison, S. (1999). Native and alien species diversity at the local and regionalscales in a grazed California grassland. Oecologia 121, 99–106. doi: 10.1007/s004420050910
He, K. S., Bradley, B. A., Cord, A. F., Rocchini, D., Tuanmu, M. N., Schmidtlein, S.,et al. (2015). Will remote sensing shape the next generation of speciesdistribution models? Remote Sens. Ecol. Conserv. 1, 4–18. doi: 10.1002/rse2.7
Heinsch, F. A., Reeves, M., Votava, P., Kang, S., Milesi, C., Zhao, M., et al.(2003). GPP and NPP (MOD17A2/A3) Products NASA MODIS Land Algorithm.Missoula, MT: University of Montana, 1–57.
Henderson, E. B., Ohmann, J. L., Gregory, M. J., Roberts, H. M., and Zald, H.(2014). Species distribution modelling for plant communities: stacked singlespecies or multivariate modelling approaches? Appl. Veget. Sci. 17, 516–527.doi: 10.1111/avsc.12085
Hengl, T., de Jesus, J. M., MacMillan, R. A., Batjes, N. H., Heuvelink, G. B. M.,Ribeiro, E., et al. (2014). SoilGrids1km — Global soil information based onautomated mapping. PLoS ONE 9:e105992. doi: 10.1371/journal.pone.0105992
Hernandez, P. A., Franke, I., Herzog, S. K., Pacheco, V., Paniagua, L., Quintana,H. L., et al. (2008). Predicting species distributions in poorly-studiedlandscapes. Biodivers. Conserv. 17, 1353–1366. doi: 10.1007/s10531-007-9314-z
Hernandez, P. A., Graham, C. H., Master, L. L., and Albert, D. L. (2006). The effectof sample size and species characteristics on performance of different speciesdistribution modeling methods. Ecography 29, 773–785. doi: 10.1111/j.0906-7590.2006.04700.x
Hestir, E. L., Khanna, S., Andrew, M. E., Santos, M. J., Viers, J. H., Greenberg,J. A., et al. (2008). Identification of invasive vegetation using hyperspectralremote sensing in the California Delta ecosystem. Remote Sens. Environ. 112,4034–4047. doi: 10.1016/j.rse.2008.01.022
Hijmans, R. J., Cameron, S. E., Parra, J. L., Jones, P. G., and Jarvis, A. (2005).Very high resolution interpolated climate surfaces for global land areas. Int. J.Climatol. 25, 1965–1978. doi: 10.1002/joc.1276
Hobbs, R. J. (1989). “The nature and effects of disturbance relative to invasions,”in Biological Invasions: A Global Perspective, eds J. A. Drake and H. A. Mooney(New York, NY: John Wiley and Sons Ltd), 389–405.
Hoffman, J. D., Narumalani, S., Mishra, D. R., Merani, P., and Wilson, R. G.(2008). Predicting potential occurrence and spread of invasive plant speciesalong the North Platte River, Nebraska. Invasive Plant Sci. Manag. 1, 359–367.doi: 10.1614/IPSM-07-048.1
Hooftman, D. A. P., Oostermeijer, J. G. B., and den Nijs, J. C. M. (2006).Invasive behaviour of Lactuca serriola (Asteraceae) in the Netherlands:spatial distribution and ecological amplitude. Basic Appl. Ecol. 7, 507–519.doi: 10.1016/j.baae.2005.12.006
Huenneke, L. F., Hamburg, S. P., Koide, R., Mooney, H. A., and Vitousek, P. M.(1990). Effects of soil resources on plant invasion and community structurein Californian serpentine grassland. Ecology 71, 478–491. doi: 10.2307/1940302
Kearney, M., and Porter, W. (2009). Mechanistic niche modelling: combiningphysiological and spatial data to predict species’ ranges. Ecol. Lett. 12, 334–350.doi: 10.1111/j.1461-0248.2008.01277.x
Kulhanek, S. A., Leung, B., and Ricciardi, A. (2011). Using ecological niche modelsto predict the abundance and impact of invasive species: application to thecommon carp. Ecol. Appl. 21, 203–213. doi: 10.1890/09-1639.1
Frontiers in Plant Science | www.frontiersin.org 16 May 2017 | Volume 8 | Article 770
https://doi.org/10.1890/11-0114.1https://doi.org/10.1111/jbi.12225https://doi.org/10.1016/j.ecolmodel.2013.09.011https://doi.org/10.1016/j.ecolmodel.2013.09.011https://doi.org/10.2307/1939503https://doi.org/10.1046/j.1365-2745.2000.00473.xhttps://doi.org/10.1111/geb.12426https://doi.org/10.1111/j.1365-2699.2011.02659.xhttps://doi.org/10.1111/j.2006.0906-7590.04596.xhttps://doi.org/10.1111/j.2006.0906-7590.04596.xhttps://doi.org/10.1016/j.foreco.2013.07.059https://doi.org/10.1016/S1405-888X(13)72073-4https://doi.org/10.1016/S1405-888X(13)72073-4https://doi.org/10.1016/j.rse.2008.01.004https://doi.org/10.1016/j.rse.2008.01.004https://doi.org/10.3390/rs1030519https://doi.org/10.3390/rs4072057https://doi.org/10.1016/j.ecolmodel.2012.07.025https://doi.org/10.1017/CBO9780511810602https://doi.org/10.1111/j.1472-4642.2010.00641.xhttps://doi.org/10.1111/j.1472-4642.2010.00641.xhttps://doi.org/10.1007/s10530-004-9642-9https://doi.org/10.1029/2011jg001733https://doi.org/10.1007/s10531-016-1206-7https://doi.org/10.1017/CBO9780511735882https://doi.org/10.1111/j.1461-0248.2005.00792.xhttps://doi.org/10.1111/j.1461-0248.2005.00792.xhttps://doi.org/10.1016/S0304-3800(00)00354-9https://doi.org/10.1007/s004420050910https://doi.org/10.1007/s004420050910https://doi.org/10.1002/rse2.7https://doi.org/10.1111/avsc.12085https://doi.org/10.1371/journal.pone.0105992https://doi.org/10.1007/s10531-007-9314-zhttps://doi.org/10.1111/j.0906-7590.2006.04700.xhttps://doi.org/10.1111/j.0906-7590.2006.04700.xhttps://doi.org/10.1016/j.rse.2008.01.022https://doi.org/10.1002/joc.1276https://doi.org/10.1614/IPSM-07-048.1https://doi.org/10.1016/j.baae.2005.12.006https://doi.org/10.2307/1940302https://doi.org/10.2307/1940302https://doi.org/10.1111/j.1461-0248.2008.01277.xhttps://doi.org/10.1890/09-1639.1http://www.frontiersin.org/Plant_Science/http://www.frontiersin.org/http://www.frontiersin.org/Plant_Science/archive
fpls-08-00770 May 12, 2017 Time: 17:38 # 17
Truong et al. Weed Risk Mapping in Southeast Asia
Le, B., Nguyen, T., and Adkins, S. (2012). Damage caused by Merremia eberhardtiiand Merremia boisiana to biodiversity of Da Nang City, Vietnam. Pak. J. WeedSci. Res. 18, 895–905.
Lozon, J. D., and MacIsaac, H. J. (1997). Biological invasions: are they dependenton disturbance? Environ. Rev. 5, 131–144. doi: 10.1139/a97-007
MacKinnon, J. R. (2002). Invasive alien species in Southeast Asia. Asean Biodivers.2, 9–11.
Matthews, S., and Brand, K. (2004). Tropical Asia Invaded: The Growing Danger ofInvasive Alien Species. Nairobi: GISP Secretariat.
McIntyre, S., Lavorel, S., and Tremont, R. (1995). Plant life-history attributes:their relationship to disturbance response in herbaceous vegetation. J. Ecol. 83,31–44. doi: 10.2307/2261148
Merow, C., Smith, M. J., and Silander, J. A. (2013). A practical guide toMaxEnt for modeling species’ distributions: what it does, and why inputsand settings matter. Ecography 36, 1058–1069. doi: 10.1111/j.1600-0587.2013.07872.x
Ministry of Natural Resources and Environment and Ministry of Agriculture andRural development (2013). Inter-Ministerial Circulation No. 27/TTLT-BTNMT-BNNPTNT. Vientiane: Ministry of Natural Resources and Environment.
Morán-Ordóñez, A., Suárez-Seoane, S., Elith, J., Calvo, L., and de Luis, E. (2012).Satellite surface reflectance improves habitat distribution mapping: a case studyon heath and shrub formations in the Cantabrian Mountains